Calibration

The calibration package optimizes BAM model parameters via a multi-phase pipeline: sensitivity analysis → focused grid screening → tiered stability testing.

Quick Start

CLI:

# Run full pipeline (Morris → grid → stability)
python -m calibration --scenario baseline --workers 10

# Individual phases
python -m calibration --phase sensitivity --scenario baseline
python -m calibration --phase grid --scenario baseline
python -m calibration --phase stability --scenario baseline

Python API:

from calibration import (
    run_morris_screening,
    build_focused_grid,
    run_focused_calibration,
    export_best_config,
)

# Phase 1: Morris screening
morris = run_morris_screening(scenario="baseline", n_workers=10, n_seeds=3)
sensitivity = morris.to_sensitivity_result()

# Phase 2: Build focused grid
grid, fixed = build_focused_grid(sensitivity)

# Phases 3-4: Screening + tiered stability
results = run_focused_calibration(grid, fixed, rank_by="combined")

# Export best config
export_best_config(results[0], "baseline")

Calibration Process

┌──────────────────┐     ┌──────────────────┐     ┌──────────────────┐
│ Phase 1          │     │ Phase 2          │     │ Phase 3          │
│ Sensitivity      │────>│ Grid Screening   │────>│ Tiered Stability │
│ (Morris or OAT)  │     │ (single seed)    │     │ (multi-seed)     │
└──────────────────┘     └──────────────────┘     └──────────────────┘
 Identifies important      Evaluates all           Tournament-style
 parameters, fixes         combinations,           multi-seed testing
 unimportant ones          prunes poor values       with ranking

Scenarios

Scenario	Parameters	Description
baseline	14 common	Standard BAM model (Section 3.9.1)
growth_plus	14 + 2 R&D	Endogenous R&D growth (Section 3.9.2)
buffer_stock	14 + 3	Buffer-stock consumption (Section 3.9.4)

Calibration Guide

Morris vs OAT: Morris Method catches interaction effects that single-baseline OAT misses. Mean sigma/mu* ratio of ~1.58 indicates pervasive interactions in the BAM model; Morris is strongly recommended for production calibration.

Parameter importance: Economy-wide params (beta, max_M) are consistently the most important. Initial conditions are largely irrelevant; the Kalecki attractor erases them in ~50 periods.

Multi-seed stability: Single-seed screening overfits to the specific random draw. A config ranking 1st with seed=0 may rank 50th across 100 seeds. Use at least 10 seeds for tier 1 and 100+ for final selection.

Multi-pass workflow: Run sensitivity → grid → stability for primary parameters, fix them at optimal values, then re-run the pipeline for previously fixed parameters. See Calibration Recipe for the full recommended workflow.